Development of an homologous recombination deficiency scoring algorithm for unmatched ovarian tumor sample based on next-generation sequencing

Abstract

Objective

Poly (ADP-ribose) polymerase inhibitors (PARPis) have emerged as a novel targeted therapy, necessitating the development of PARPi sensitivity tests for clinical application. However, the requirement for matched normal tissue complicates clinical testing procedures and increases the burden on patients. In this study, we optimized a bioinformatics algorithm utilizing the Panel of Normals to accommodate tumor-only scenarios.

Methods

Paired samples from both clinical (Chinese population) and International Cancer Genome Consortium (ICGC) databases (Caucasian population) were employed to establish training cohorts. The tumor-only algorithm was developed based on these cohorts. The performance of the tumor-only algorithm was evaluated through linear regression against the paired sample algorithm. The validity of both algorithms was tested using progression-free survival data from patients treated with PARPis. Additionally, clinical samples from an independent institution were utilized to further validate the tumor-only algorithm.

Results

In the training sets, we observed high correlations between the tumor-only and paired sample algorithms regarding tumoral purity and homologous recombination deficiency (HRD) scores, with most R² values exceeding 0.9. In the validation set, a slight decrease in correlation was noted, although the majority of R² values remained close to 0.9. Both algorithms effectively distinguished PARPi-sensitive patients. No statistically significant differences were identified between the training and validation cohorts concerning clinical characteristics. Cross-racial validation yielded similar results.

Conclusion

The tumor-only algorithm demonstrated an equivalent capacity to the paired sample algorithm for classifying HRD status in ovarian cancer. This algorithm also showed cross-racial applicability, highlighting its potential for clinical use.

Synopsis

Matched normal tissue complicates testing procedures, and a tumor-only algorithm was developed for homologous recombination deficiency scoring. The tumor-only algorithm was validated by comparing it to the paired sample algorithm and matching clinical functional profiling system data. This new tumor-only algorithm could be used cross-racially with a very small Panel of Normals.

INTRODUCTION

Therapies based on the concept of synthetic lethality have emerged as a novel class of targeted therapies over the past decade. The underlying mechanism involves the simultaneous deficiency of 2 or more signaling pathways, which results in cellular death, whereas the dysfunction of any individual pathway does not compromise cell survival. Several drugs developed on this principle have demonstrated promising outcomes in clinical trials [1]. Among these, poly (ADP-ribose) polymerase inhibitors (PARPis), particularly olaparib, have been proposed for the treatment of patients with BRCA1/2 mutations and have received approval for clinical use from the U.S. Food and Drug Administration (FDA) [2]. Initially, olaparib was indicated for patients with ovarian cancer (OC) harboring BRCA mutations [3]. Subsequently, the FDA expanded its indication to include patients with breast cancer, pancreatic cancer, or prostate cancer [4,5,6] who present with homologous recombination repair deficiency (HRD) [7].

With the approval of additional PARPis, such as rucaparib, niraparib, and talazoparib [8,9,10], the clinical utilization of PARPis has become commonplace, leading to the establishment of companion diagnostics (CDx) as routine tests. Currently, the most frequently utilized CDx or genetic tests for PARPis are the BRCA1 and BRCA2 tests. In addition to BRCA1 and BRCA2, multiple other genes contribute to HRD, including ATM, RAD51, and various genes from the Fanconi anemia complementation group [11]. However, directly testing for mutations in these genes to indicate HRD is impractical, as the severity of HRD associated with mutations in these genes remains unknown.

Given that genomic instability is a hallmark of HRD, several alternative methods for quantifying HRD, such as the genomic instability score, have been developed. These methods include the assessment of loss of heterozygosity (LOH) [12], telomeric allelic imbalance (TAI) [13], large-scale state transitions (LST) [14], and their combination, termed the HRD score [15]. Most of these methods were originally based on single-nucleotide polymorphism (SNP) array technology. However, with advancements in sequencing technology, these array-based techniques have gradually been supplanted by next-generation sequencing (NGS), which has become the predominant method for genomic characterization of cancer biopsies. Alongside this technological evolution, bioinformatics algorithms have advanced to generate comparable outcomes from NGS data [16].

In a previous study, we introduced HiSNP-ultra, an NGS-specific SNP panel designed for HRD measurement using paired clinical samples [17]. However, the absence of matched normal tissue in clinical settings is common, as archived specimens in most clinical laboratories typically do not include normal tissues or blood. The requirement for normal tissue complicates clinical and testing procedures, as it necessitates additional informed consent, the acquisition of new specimens, and supplementary testing. This not only increases procedural complexity and economic burdens for patients but also hinders the utilization of the HiSNP-ultra panel. In this study, we further optimized the algorithm to accommodate a tumor-only scenario. The principles underlying this algorithm are applicable not only to the HiSNP-ultra panel but also to other similar HRD testing panels. Additionally, we demonstrate that this algorithm can be effectively applied cross-racially for HRD scoring in OC.

MATERIALS AND METHODS

1. Sample inclusion

Formalin-fixed paraffin-embedded (FFPE) tumoral specimens with matched peripheral blood samples were collected from patients with OC. A total of 81 paired samples were obtained from Harbin Cancer Hospital between January 1, 2013, and December 30, 2022. Of the corresponding 74 patients from this cohort, progression-free survival (PFS) data were available (Tables S1 and S2). An additional 34 paired samples were collected from Henan Cancer Hospital between January 1, 2017, and December 30, 2023. All patients were clinically classified as either platinum-sensitive (defined as no recurrence or recurrence occurring >6 months after initial therapy) or platinum-resistant (recurrence within ≤6 months of initial therapy). PARPi (Olaparib, Niraparib and Fluzoparib) was administered as first-line therapy. All patients’ pathological classifications and treatment are listed in Tables S1 and S2. Furthermore, whole-genome sequencing (WGS) data for 70 ovarian carcinomas with paired normal controls were downloaded from the International Cancer Genome Consortium (ICGC) database for use as a simulated validation dataset. The study protocol was approved by the Ethics Committees of Harbin Cancer Hospital (KY2021-37) and Henan Cancer Hospital (2023-KY-0059). This study was conducted in accordance with the ethical principles of the 1964 Helsinki Declaration and its later amendments. Written informed consent was obtained from all participants.

2. DNA extraction

Tumoral specimens were processed into paraffin blocks following standard pathological procedures at the respective hospitals. For each sample, one 4 μm section was stained with hematoxylin and eosin for histological assessment. Genomic DNA was subsequently extracted from two to five 10 μm sections using the QIAamp DNA FFPE Tissue Kit (Cat. #56404; Qiagen, Hilden, Germany), according to the manufacturer’s instructions. DNA concentration was quantified using a Qubit™ 4 Fluorometer (Cat. # Q33226; Thermo Fisher Scientific, Waltham, MA, USA).

3. NGS and analysis pipeline

DNA libraries were prepared from 50–100 ng of input DNA using the NadPrep Universal DNA Library Preparation Kit (Cat. #1002212; Nanodigmbio, Nanjing, China). Fragmentation was performed using a Covaris™ M220 Focused-ultrasonicator (Cat. #500295; Covaris, Woburn, MA, USA). Following end repair, A-tailing, adapter ligation, and clean-up steps, ligated fragments were amplified for 8–12 cycles using a 0.5 M index primer mix. Final library yields were adjusted to 500–1,000 ng.

Hybrid capture was performed using NadPrep Hybrid Capture Reagents (Cat. #1005101; Nanodigmbio). A custom 8.6 Mb hybrid capture panel was constructed, comprising the NanOnco Plus Panel v3.0 (2.4 Mb) and the HiSNP Ultra Panel v1.0 (6.2 Mb). The NanOnco Plus Panel targets the complete coding regions of 620 genes relevant to pan-cancer studies, along with intronic sequences associated with common gene fusions. The HiSNP Ultra Panel targets over 52,000 high minor allele frequency (MAF) SNPs distributed evenly across the human genome at ~50 kb intervals.

The 81 paired samples from Harbin Cancer Hospital were captured using the HiSNP Ultra Panel v1.0 alone, while the 34 paired samples from Henan Cancer Hospital were captured using the full 8.6 Mb hybrid panel. Post-capture amplification was conducted via polymerase chain reaction for 13–15 cycles. Libraries were sequenced in a 150 bp paired-end mode on a DNBSEQ-T7 sequencer (MGI Tech, Shenzhen, China), achieving an average sequencing depth of 500×.

Raw sequencing reads were processed by removing adapters and low-quality sequences. Clean reads were aligned to the human reference genome (GRCh37) using BWA (version 0.7.17). Duplicate reads were marked and removed using the MarkDuplicates tool in Picard (version 4.0.4.0). Local realignment around putative indels and single nucleotide variants (SNVs), along with base quality score recalibration, was performed using Genome Analysis Toolkit (GATK, version 3.8.0). Tumoral purity and ploidy were estimated from the aligned data using the Sequenza R package. Allele-specific copy number (ASCN) analysis was subsequently performed based on these estimates. For more detailed methods, please refer to Data S1.

4. ICGC WGS dataset simulation

The ICGC database was queried for "Ovarian cancer" projects meeting the following criteria: WGS data with a minimum depth of 30× and availability of matched tumor-normal pairs. Seventy such WGS datasets were downloaded. Reads mapping to the genomic regions covered by the HiSNP Ultra Panel were extracted in silico and analyzed using the identical bioinformatics pipeline applied to the clinical samples.

5. Copy number variation profiling in tumor-only samples

Allele-specific coverage at approximately 52,000 high-MAF SNP loci within the HiSNP Ultra Panel was compared between each tumor sample and a Panel of Normals (PoN) samples. Heterozygous SNP sites were identified using 3 independent callers: GATK HaplotypeCaller, VarDict, and VarScan; coverage depth and B-allele frequency (BAF) were compared across callers. For copy number (CN) profiling, an in-house CNV calling pipeline (“depth ratio”) was employed. Briefly, the target region was scanned using a 120-bp sliding window. The normalized coverage in the tumor was divided by the normalized average coverage from the PoN to compute a log₂ ratio for each window. A least-squares segmentation algorithm was applied to define CN segments. Within a given region, original log₂ values were retained unless classified as outliers (exceeding 3 standard deviations from the local mean), in which case they were winsorized. Segmentation was performed by minimizing the objective function:

$$L\left(S\right|y, z) = \sum _{I\in S}\sum _{j\in I}{(yj - \stackrel{-}{y}I)}^2 + z * |S|$$

Where L is the least square in the region S; yj is the adjusted log₂ values and ȳI is the mean of yj, I is collection of sequential probes share same log₂ value, i.e., I1 consists of probes with same log₂ value until hit the first breakpoint, and I2 consists of the subsequent probes until the next breakpoint, and so on. The specific regional S = {I1, I2..., IM}. |S| is the number of segments in S, and z is a positive constant number that controls the measurement noise. Then ASCN was estimated from SNP and CNV profiles by using the Sequenza R package.

6. HRD score calculation

Tumor sample HRD score was determined using an algorithm similar to that described by Popova et al. [14]. TAI (number of regions of allelic imbalance that extend to one of the sub-telomeres but do not cross the centromere) and LST (number of breakpoints between regions longer than 10 Mb after filtering out regions shorter than 3 Mb) scores were calculated using the allelic imbalance profiles, while LOH (number of sub-chromosomal LOH regions longer than 15 Mb) was calculated using ASCN. Homologous recombination deficiency (HRD) score was defined as the sum of TAI, LST, and LOH scores. Sequenza preprocessed files and other allele-specific segmentation files in the same format were treated using the scarHRD R package (https://github.com/sztup/scarHRD) to calculate the HRD score.

7. PFS curve

The start date of PFS was counted at the end date of the last PARPi therapy. The follow-up was carried out every 3 months until recurrence happened.

8. Statistical analysis

Statistical analysis was performed with SPSS v24.0 for Windows (SPSS, Chicago, IL, USA). The Student’s t-test was used for continuous variables (i.e., age) and the Fisher’s exact test for categorical variables (i.e., clinical histology). The p-value <0.05 was considered a significant difference.

RESULTS

1. Characteristics of samples and data

A total of 78 tumor-normal paired clinical specimens from OC patients were collected from Harbin Cancer Hospital for the development of the tumor-only algorithm. The ages of the patients ranged from 39 to 70 years, with a median age of 54 years. The majority of patients presented at International Federation of Gynecology and Obstetrics (FIGO) stage III (66 out of 81, 81.5%), and the pathological diagnosis predominantly consisted of high-grade serous carcinoma (72 out of 81, 88.9%). Additionally, 70 pairs of WGS data from OC patients were obtained from the ICGC-OV database [18] to facilitate the simulated study. An extra 34 paired clinical samples were sourced from Henan Cancer Hospital for independent validation of the algorithm. The characteristics of the 2 clinical cohorts were compared, and statistical assessments were conducted to determine differences in clinical features. No significant differences were observed between the cohorts regarding age, FIGO stage, BRCA1/2 mutations, or recurrence rates (Table S3). Continuous variables, such as age, were analyzed using Student’s t-test, while categorical variables, such as clinical histology, were assessed using Fisher’s exact test. A p-value of less than 0.05 was considered indicative of a statistically significant difference.

2. Schematic workflow of bioinformatic analysis

The 81 pairs of OC samples from Harbin Cancer Hospital (referred to as Chinese cohort and hereafter) were target sequenced using the HiSNP-ultra panel. The normal samples in the cohort were used as the PoN and variants called from the PoN were used to create the baseline. Loci of the HiSNP-ultra panel were extracted from 70 pairs of ICGC-OV dataset (referred to as Caucasian cohort and hereafter) and the same approach was applied to create another parallel baseline. The baselines were used as the reference for the bioinformatic calculations in the tumor-only algorithm. To assess the performance of the tumor-only algorithm, the algorithm was compared with the standard paired algorithm regarding the outcomes of tumoral purities and HRD scores using the same samples in the Chinese cohort. (Fig. 1). Meanwhile, a synchronous simulation was also conducted using the same method in the Caucasian cohort for further evaluation (Step 1). After the initial validation, the size of PoN was down-scaled to various levels to find out the minimum number for creating the PoN (Step 2). The cross-racial performance of the tumor-only algorithm was also tested by swapping PoNs between the Chinese cohort and the Caucasian cohort (Step 3). Lastly, a new independent clinical cohort was tested to validate the algorithm (Step 4).

Fig. 1. Schematic diagram of the workflow for developing the tumor-only HRD scoring algorithm.

HRD, homologous recombination deficiency; ICGC, International Cancer Genome Consortium; WGS, whole-genome sequencing.

3. Tumoral purity validation

Accurate quantification of tumoral content is critical for the precise calculation of copy number variation (CNV). To assess the consistency of tumoral purity estimates derived from both tumor-only and paired algorithms, we first compared the outcomes using the same samples. We evaluated our in-house CNV calling method (depth ratio) in conjunction with 3 SNV calling methods: GATK HaplotypeCaller [19], VarDict [20], and VarScan [21]. This evaluation aimed to identify the optimal algorithmic combination for input into Sequenza [22,23,24], which serves as a prerequisite for HRD scoring via ScarHRD [16]. Subsequently, ScarHRD was employed for HRD computation, and tumoral purities were extracted from the generated results.

To quantify the correlation between the outcomes of the 2 algorithms, we applied standard linear regression analysis. Across all combinations of variant calling tools, we observed high consistency between the tumor-only and paired algorithms in both the Chinese cohort (clinical data) and the Caucasian cohort (simulated data), yielding R² values of 0.9193, 0.8633, and 0.9145 for VarScan, VarDict, and GATK in the Chinese cohort, respectively. In the Caucasian cohort, the corresponding R² values were 0.9224, 0.9371, and 0.8538 (Fig. S1). Notably, VarScan exhibited robust performance across both cohorts (Fig. 2A and C), while VarDict demonstrated superior results in the Caucasian cohort (Fig. S1B and E), and GATK HaplotypeCaller performed better in the Chinese cohort (Fig. S1C and F). Consequently, we selected VarScan as our preferred SNV caller for subsequent analyses. These findings underscore the high reliability of the tumor-only algorithm in calculating tumoral purity.

Fig. 2. Initial assessment of the tumor-only HRD scoring algorithm. Comparison of tumoral purities (A) and the HRD scores (B) between the tumor-only and the paired algorithm from the same clinical samples. And the comparison of tumoral purities (C) and the HRD scores (D) between the tumor-only and the paired algorithm from the same International Cancer Genome Consortium samples (grey shadows indicate confidence intervals). Calculated copy number variations by the tumor-only algorithm (E) and the paired algorithm (F) in clinical samples. Grey dotted lines indicate 30% tumoral purity (A, C); Black dashed lines indicate the 42 thresholds for the differencing HRD deficiency and proficiency (B, D); Green and red dots indicate consistent cases and grey dots indicate inconsistent cases between the 2 algorithms (B, D).

HRD, homologous recombination deficiency.

4. HRD scores with or without tumoral purity filtration

We proceeded to compare the HRD scores computed by the tumor-only and paired algorithms using the same combinations of SNV calling tools and assessment methods employed in the tumoral purity analysis. Initially, we assessed the correlations of HRD scores without applying any tumoral purity filtration. The results indicated poor correlations between the 2 algorithms in the Chinese cohort (Fig. S2A-C, blue dashed lines), whereas a strong correlation was evident in the Caucasian cohort (Fig. S2D-F, blue dashed lines). Current FDA-approved HRD tests mandate a minimum tumoral content of 30% or 35%. Based on our prior observations [17], we opted to implement a 30% tumoral purity threshold for sample filtration. Following this filtration, the consistency between the tumor-only and paired algorithms significantly improved (Fig. S2, blue solid lines). A representative result (VarScan combined with depth ratio) is illustrated in Fig. 2B and D. Importantly, approximately half (42 out of 81) of the clinical samples were retained post-filtration, with only one sample from the ICGC cohort excluded.

Given that CNVs underpin HRD score calculations, we further examined CNV patterns generated by both the tumor-only and paired algorithms in the Chinese cohort after filtration. Our analysis revealed that CNV patterns were nearly indistinguishable between the 2 algorithms (Fig. 2E and F). Similar results were observed in the Caucasian cohort (Fig. S3). Collectively, these results suggest that the HRD scores obtained from the tumor-only and paired algorithms are almost equivalent following the application of a 30% tumoral purity filtration.

5. Reduction of PoN size

The use of a PoN is an effective strategy for analyzing tumor-only samples. In many instances, even a small PoN can provide advantages over the absence of a PoN entirely. However, no definitive guidelines exist regarding the optimal number of samples to be included in a PoN, given the considerable variability among NGS pipelines. The GATK recommends a minimum of 40 samples. In our previous work, we utilized all 81 normal controls to construct the PoN. Nonetheless, many laboratories may face challenges in assembling a large PoN due to difficulties in obtaining sufficient normal tissues. Therefore, we aimed to investigate how variations in PoN size impact the performance of the tumor-only algorithm.

Initially, we reduced the PoN size to 40 samples and evaluated the consistency between the tumor-only and paired algorithms. A strong correlation was observed between the 2 methods in both the Chinese and Caucasian cohorts (Fig. 3A-D). We focused on the VarScan and depth ratio combination, as this pairing demonstrated optimal performance in prior assessments. Subsequently, we conducted a comprehensive trial involving systematic downsizing of the PoN from 40 to 1 sample, with and without tumoral purity filtration. When the PoN size was minimized to one, we repeated the process 5 times randomly and averaged the results. As illustrated in Fig. 3E and F, the assessment of tumoral purity appeared unaffected by the size of the PoN in both cohorts, with R² values exceeding 0.9, even when the PoN was reduced to a single sample. Conversely, when tumoral purity was not applied, HRD scores calculated using the tumor-only algorithm were poorly correlated within the Chinese cohort; however, a significant improvement in correlation was observed upon applying purity filtration (Fig. 3E). Notably, the consistency of HRD scores remained unaffected by tumoral purity in the Caucasian cohort (Fig. 3F). When utilizing the minimal PoN (one normal sample), the tumor-only algorithm displayed inferior performance regarding HRD scores in the Chinese cohort, with an R² value of less than 0.9 (Fig. 3E, right cluster). In contrast, the HRD scores consistently correlated well between the tumor-only and paired algorithms, exhibiting low variance with the minimal PoN (Fig. 3F, right cluster). These findings suggest that a very small PoN is sufficient for the tumor-only algorithm to function effectively.

Fig. 3. The assessment of downsizing PoN. Comparison of tumoral purities between the tumor-only and the paired algorithm with the downsizing PoN of 40 in the clinical samples (A) and ICGC samples (B). And the comparison of HRD scores between the tumor-only and the paired algorithm with the downsizing PoN of 40 in the clinical samples (C) and ICGC samples (D). Shadows in panel (A-D) indicate confidence intervals. Observed R² between the tumor-only and the paired algorithm when the PoN was downsized to 40, 20, 10, 5, and 1 in the clinical samples (E) and ICGC samples (F). Grey dotted lines indicate 30% tumoral purity (A, B); Black dashed lines indicate the 42 thresholds for the differencing HRD deficiency and proficiency (C, D); Green and red dots indicate consistent cases and grey dots indicate inconsistent cases between the 2 algorithms (C, D); Error bar was the standard deviation and the corresponding bar was the average result of 5 random normal control (E, F).

HRD, homologous recombination deficiency; ICGC, International Cancer Genome Consortium; PoN, Panel of Normals.

6. Cross-racial validation

Thus far we have evaluated the tumor-only algorithm for HRD measurement using the HiSNP-ultra panel across both Chinese and Caucasian cohorts, demonstrating comparable results to the paired algorithm. A common practice among many laboratories involves leveraging public data sources to create PoNs without accounting for ethnic background. To address this issue, we tested the tumor-only algorithm using cross-racial PoNs. In this evaluation, PoNs derived from normal samples of the Chinese cohort were employed to calculate tumoral purities and HRD scores in the Caucasian cohort, and vice versa. We also examined the effects of serial downsizing of the PoN.

Surprisingly, this cross-racial PoN performed well in calculating HRD scores, yielding R² values of 0.8987 and 0.9424 for the Chinese and Caucasian cohorts, respectively (Fig. 4A and B). Even with reductions in PoN size, the tumor-only algorithm maintained effective performance, albeit with slightly diminished efficacy compared to the original racially matched PoNs (Fig. 4C and D). Again, the algorithm struggled without tumoral purity filtration in the Chinese cohort, and the minimal PoN of one sample exhibited inferior performance compared to larger PoNs.

Fig. 4. Cross-racial use of the PoNs. The PoNs were cross-racially used. The comparison of HRD scores between the tumor-only and the paired algorithm was shown in Chinese patients with the Caucasian PoN (A) or in Caucasian patients with the Chinese PoN (B); Dashed lines indicate the threshold of 42 for the differencing HRD deficiency and proficiency; Green and red dots indicate consistent cases and grey dots indicate inconsistent cases between the 2 algorithms. Observed R² between the tumor-only and the paired algorithm in the PoN swapping setting when the PoN was downsized to 40, 20, 10, 5, and 1 in the clinical samples (C) and ICGC samples (D); The error bar was the standard deviation and the corresponding bar was the average result of 5 random normal controls. When assuming the paired algorithm was the ground truth, the calculated area under the curve by the tumor-only algorithm for the clinical samples (E) and the ICGC samples (F). Shadows in panel (A, B, E, F) indicate confidence intervals.

HRD, homologous recombination deficiency; ICGC, International Cancer Genome Consortium; PoN, Panel of Normals; ROC, receiver operating characteristic.

To assess the overall effectiveness of the tumor-only algorithm, we computed the area under the receiver operating characteristic curve (AUROC) for both cohorts. We assumed that HRD scores derived from paired samples represented the true values and established a cutoff of 42 for classifying samples as HRD proficient or HRD deficient. By comparing classifications between the tumor-only and paired algorithms, we were able to determine metrics such as true positives, false positives, true negatives, and false negatives, facilitating the calculation of sensitivities and specificities. As depicted in Fig. 4E and F, the tumor-only algorithm demonstrated robust performance in both the Chinese and Caucasian cohorts. In the Chinese cohort, classifications from the 2 algorithms were identical when filtration was applied, while the AUROC approached one in the Caucasian cohort. Moreover, the tumor-only algorithm yielded results comparable to those of the paired algorithm within this cross-racial PoN framework.

7. HRD threshold for Chinese patients

While a cutoff value of 42 has been utilized in previous literature and in the tests conducted in this study, the appropriate threshold for HRD scoring specific to Chinese patients remains undetermined and is essential for the clinical validation of this novel algorithm. Given that PFS is one of the most commonly used indicators for assessing drug efficacy, and considering that HRD scores serve as guidance for PARPis, we employed PFS as a benchmark to identify the optimal threshold for the HiSNP-ultra panel in tumor-only mode.

To this end, we examined various cutoff values ranging from 35 to 45 to categorize patients into sensitive and resistant groups. We then analyzed the differences in PFS between these 2 cohorts, with greater differences indicating a more suitable cutoff. As illustrated in Fig. 5A, varying cutoff values in combination with different PoNs yielded a range of p-values, reflecting the significance of PFS differences between the 2 patient groups. The overall trend observed was a decrease in p-values starting from the cutoff value of 35, reaching a plateau around 41. The typical PFS curves for the 2 patient groups classified by an HRD score of 41 are displayed in Fig. 5B. In all tested conditions, including both paired and tumor-only algorithms, the 2 classified groups exhibited significant differences, with p-values less than 0.05, thereby supporting the appropriateness of the selected cutoff value. Notably, patients demonstrating improved PFS were distinctly separated from those with poorer outcomes, indicating that scoring with HiSNP-ultra aligns well with actual clinical results. Additional PFS curves corresponding to cutoff value screening are provided in Fig. S4.

Fig. 5. HRD cutoff value determination for Chinese patients. HRD cutoff values from 35 to 45 were selected and patients were classified into sensitive and resistant groups correspondingly. The differences in PFS between those 2 groups were measured by p-values. A lower p-value indicates better separation between 2 groups. The overall outcomes of all p-values generated using various cutoff values and Panel of Normals combinations are presented (A). The typical PFS curves of 2 patient groups that were classified by a cutoff value of 41 are displayed (B).

HRD, homologous recombination deficiency; PFS, progression-free survival.

We also conducted a similar cutoff screening analysis using data without the 30% tumoral purity filtration; while the p-values exhibited a comparable trend, they were more scattered (Fig. S5).

8. Independent clinical validation

To assess the robustness of the tumor-only algorithm, we examined 34 independent paired OC samples from Henan Cancer Hospital. All patients included in this study had received platinum-based therapy and were classified as either platinum-sensitive or platinum-resistant based on the occurrence of recurrence within 6 months following the initial chemotherapy treatment. We utilized the previously generated PoN from the Chinese cohort to calculate tumoral purities and HRD scores, subsequently evaluating the concordance between the tumor-only and paired algorithms.

The estimation of tumoral purity demonstrated strong concordance, with an R² of 0.8498 (Fig. 6A). For HRD scores, 21 samples remained after filtration based on the 30% tumoral purity criteria, and these scores also showed good agreement, indicated by an R² of 0.8939 (Fig. 6B). To compare the performance of the tumor-only algorithm to the standard method [15], we selected an additional fifteen clinical samples that had both NGS and chromatin immunoprecipitation (CHIP) assay data. We calculated HRD scores using both the tumor-only algorithm and the CHIP-based standard method. The results showed that HRD scores from the 2 different methods were closely aligned, with an R² of 0.84402 (Fig. 6C), indicating the validity of the tumor-only algorithm. The calculated HRD scores derived from tumor-only and paired algorithms are presented in Fig. 6D. As shown, the majority of HRD scores were highly comparable between the paired and tumor-only methods (illustrated by similar bar heights in Fig. 6D). We further assessed the difference in HRD scores between the platinum-sensitive and resistant groups, finding significant distinctions using both algorithms (Fig. 6E and F). Additionally, the AUROC demonstrated a well-covered area of 0.92, indicating strong consistency between the paired and tumor-only algorithms (Fig. S6). Similar findings were obtained when utilizing the Caucasian PoN (Fig. S6).

Fig. 6. Independent clinical validation. Thirty-four independent paired samples from the outside of the institute were tested. The tumoral purities were compared between the tumor-only and the paired algorithm using the Chinese PoN (A). Grey dotted lines indicate 30% tumoral purity. The HRD scores were also compared in the same way with the Chinese PoN (B). Dashed lines indicate the threshold of 41 for the differencing HRD deficiency and proficiency; Green and red dots indicate consistent cases and grey dots indicate inconsistent cases between the 2 algorithms. A side-by-side comparison of HRD scores between the paired and tumor-only algorithms in 30% tumoral purity filtered samples (n=21) is exhibited; (C). Shadows in panel (A, B, C) indicate confidence intervals. Comparison of HRD scores between tumor-only algorism and standard CHIP based method; (D). Distributions of HRD scores in the same set of clinically defined platinum-sensitive and platinum-resistant patients (E); HRD scores calculated using paired algorithm; (F); HRD scores calculated using the tumor-only algorithm.

HRD, homologous recombination deficiency; PoN, Panel of Normals.

DISCUSSION

In this study, we collected 81 paired clinical OC samples from Harbin Cancer Hospital, 70 paired OC data from the ICGC, and an additional 34 paired clinical OC samples from Henan Cancer Hospital. The aim was to develop and validate a tumor-only algorithm for HRD scoring utilizing the HiSNP Ultra Panel in OC. Rather than relying on matched control samples, the tumor-only algorithm employed a PoN for all samples. We calculated tumoral purities and HRD scores using both the tumor-only and paired algorithms from the same samples, assessing the consistency between the 2 approaches. Our findings indicated that the overall results from the tumor-only algorithm were comparable to those of the paired algorithm. Given that matched normal tissue is often unavailable in clinical settings, this tumor-only algorithm significantly broadens the applicability of the HiSNP Ultra Panel and similar SNP backbone panels for HRD scoring.

To evaluate the robustness of the tumor-only algorithm, we utilized an independent external cohort for validation. We compared and analyzed the clinical characteristics between the training cohort (Harbin Cancer Hospital) and the validating cohort (Henan Cancer Hospital) using statistical tests. No significant differences were observed between the 2 groups concerning age, clinical stage, gene mutations, and other characteristics, indicating unbiased sample selection for the tumor-only algorithm (Table S3).

For the calculation of tumoral purity, the tumor-only algorithm demonstrated robust consistency with the conventional paired algorithm. High levels of concordance in tumoral purity were observed across all settings, including the initial training phase (Fig. 2A and C), PoN downsizing (Fig. 3A and B), and independent validation (Fig. 5A and B). We attribute this strong agreement primarily to the use of BAF for purity calculations. Since the BAF in tumors remains unchanged regardless of whether a normal control is employed, the estimated tumoral purity based on BAF should not exhibit significant variation. However, one potential influencing factor could be the method used for SNV calling, which may introduce subtle differences, as illustrated in Fig. S1.

With respect to HRD scores, we identified tumoral purity as a significant influential factor, necessitating filtration with a 30% cutoff value, consistent with our previous findings [17]. We observed that clinical samples exhibited poor consistency between the 2 algorithms before filtration, but this improved markedly afterward (Figs. 3E and 4C, Fig. S2A-C). We attribute this improvement to the fact that HRD scoring relies on CNV analysis. Because clinical specimens are often heterogeneous and contain normal diploid cells (e.g., infiltrating immune cells and stromal cells), their presence in bulk sequencing data biases CNV calculation for aneuploid tumor cells. Therefore, high tumor purity is essential for accurate results. After filtration, approximately half of the clinical samples were excluded from the dataset, while the majority of ICGC samples were preserved, with only one exclusion. This suggests that most ICGC samples were pre-selected to meet the database’s quality requirements. Consequently, this pre-selection likely explains why results from the Caucasian cohort were generally better (Fig. 3B) and less affected by tumoral purities (Fig. 3F). In real-world studies, clinical specimens are often heterogeneous and may not fulfill high-quality standards; thus, caution is warranted in developing algorithms using large public databases such as The Cancer Genome Atlas and ICGC.

One of the most surprising findings of this study is that the PoN size could be reduced to as few as one sample, and this PoN can be applied cross-racially (Figs. 3E, 3F and 4). We propose that this is due to the underlying algorithm for HRD scoring being reliant on large CNVs. Unlike SNVs, which are common and highly diverse among individuals, large CNVs are rare and typically do not occur in healthy tissues. The HRD scoring integrates LOH, LST, and TAI. LOH detects losses of fragments over 15 Mb, while LST monitors chromosomal breaks between adjacent regions of at least 10 Mb, ensuring the distance between them does not exceed 3 Mb. These events represent large CNVs that are unlikely to occur in healthy tissue, irrespective of racial or ethnic background. Therefore, the differences in large CNVs are minimal among individuals, whether from the same or different racial/ethnic groups. This explains why the PoN can be minimized to just one person and still maintain effective performance across diverse populations. As for the consistency of tumoral purity and HRD scores, the tumor-only algorithm showed slightly diminished performance in the validating cohort (Fig. 6A-D). Nonetheless, the overall results remained favorable for classifying samples as HRD deficient or HRD proficient (Fig. 6E and F). We believe this observation is due to the score-based nature of the classification; minor differences in scores will not affect outcomes unless they are close to the cutoff threshold. Traditionally, the HRD score was calculated using the CHIP assay and has been accepted as a standard measure of HRD status [15]. When we compared the HRD scores obtained from the standard method with those from the tumor-only algorithm, we found them to be very similar (Fig. 6C). These findings further indicate the applicability of the tumor-only algorithm for real clinical samples.

In the past decade, the treatment landscape for OC has rapidly evolved beyond traditional chemotherapy, with novel targeted agents demonstrating significant promise in improving outcomes for specific patient subgroups. Olaparib, niraparib, and rucaparib are now established as standard maintenance therapies for advanced high-grade serous ovarian cancer in patients with HRD, particularly among those with BRCA mutations [25].

We developed a tumor-only algorithm for HRD scoring utilizing the HiSNP Ultra Panel. This algorithm necessitates the use of a PoN and requires a minimum of 30% tumoral purity in the specimens. While the PoN can be minimized to a single normal sample and applied cross-racially, we recommend using at least 10 normal samples to establish a reliable PoN baseline, with a suggested cutoff value of 41. The tumor-only algorithm demonstrated robust consistency with the conventional paired algorithm and was validated using clinical functional profiling system data. While this algorithm currently serves as a reliable alternative when matched normal samples are unavailable, it could become a full replacement for paired testing with further validation.

This research is retrospective in nature and is limited by the total sample size. A prospective study with a larger sample size is warranted to further validate these findings.

SUPPLEMENTARY MATERIALS

Data S1

Methods

Table S1

Diagnose and PFS of patients from Harbin Cancer Hospital

Table S2

Diagnose and clinical information of patients from Henan Cancer Hospital

Table S3

Characteristics of patients from 2 medical centers

Fig. S1

Comparison of tumoral purities between the tumor-only and the paired algorithm using different SNV calling tools. The in-house developed bioinformatical pipeline (depth ratio) in combination with open-source SNV calling software Genome Analysis Toolkit haplotype caller, Vardict, and Varscan were used for the tumor-only tumoral purity calculation. The y-axis is the purity calculated via the tumor-only algorithm; the x-axis is the purity calculated via the paired sample algorithm. The top row (A-C) was the result of the Chinese cohort (clinical data); the bottom row (D-F) was the result of the Caucasian cohort (simulated data). The correlation between the tumor-only and paired algorithms was measured by a standard linear regression. Grey dotted lines indicate 30% tumoral purity. Shadows indicate confidence intervals.

Fig. S2

Comparison of HRD scores between the tumor-only and the paired algorithm using different SNV and copy number variation calling tools and with or without tumoral purity filtration. The in-house developed bioinformatical pipeline (depth ratio) in combination with open-source SNV calling software Genome Analysis Toolkit haplotype caller, Vardict, and Varscan were used for the tumor-only HRD measurement. The y-axis is the HRD score calculated via the tumor-only algorithm; the x-axis is the HRD score calculated via the paired sample algorithm. The top row (A-C) was the result of the Chinese cohort (clinical data); the bottom row (D-F) was the result of the Caucasian cohort (simulated data). The correlation between the tumor-only and paired algorithms was measured by a standard linear regression. The blue dashed and solid lines indicate the regression lines before and after the filtration respectively. Black dashed lines indicate the 42 thresholds for the HRD score; Green and red dots indicate consistent cases and grey dots indicate inconsistent cases between the 2 algorithms. Shadows indicate confidence intervals.

Fig. S3

Calculated copy number variations by the tumor-only algorithm (A) and the paired algorithm (B) in International Cancer Genome Consortium samples.

Fig. S4

Additional PFS curves of 2 patient groups that were classified by a cutoff value of 41.

Fig. S5

The p-values generated using various cutoff values and Panel of Normals combinations in samples without 30% tumoral purity filtration. HRD cutoff values from 35 to 45 were selected and patients were classified into sensitive and resistant groups correspondingly. The differences in PFS between those 2 groups were measured by p-values. A lower p-value indicates better separation between 2 groups. The overall outcomes of all are presented.

Fig. S6

Thirty-four independent paired samples from the outside of the institute were tested. The tumoral purities were compared between the tumor-only and the paired algorithm using the Caucasian PoN (A); Grey dotted lines indicate 30% tumoral purity. The HRD scores were also compared in the same way with the Caucasian PoN (B); Dashed lines indicate the threshold of 41 for the differencing HRD deficiency and proficiency; Green and red dots indicate consistent cases and grey dots indicate inconsistent cases between the 2 algorithms. The area under the curve was calculated for the tumor-only algorithm assuming the paired algorithm was the ground truth when the Chinese PoN (C) and the Caucasian PoN (D) were used. Shadows indicate confidence intervals.